Re: [PATCH 31/35] host/include/aarch64: Implement aes-round.h

[Richard Henderson <richard.henderson@linaro.org>]

On 6/3/23 05:50, Ard Biesheuvel wrote:
> On Sat, 3 Jun 2023 at 04:34, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>>
>> Detect AES in cpuinfo; implement the accel hooks.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>>   host/include/aarch64/host/aes-round.h | 204 ++++++++++++++++++++++++++
>>   host/include/aarch64/host/cpuinfo.h   |   1 +
>>   util/cpuinfo-aarch64.c                |   2 +
>>   3 files changed, 207 insertions(+)
>>   create mode 100644 host/include/aarch64/host/aes-round.h
>>
>> diff --git a/host/include/aarch64/host/aes-round.h b/host/include/aarch64/host/aes-round.h
>> new file mode 100644
>> index 0000000000..27ca823db6
>> --- /dev/null
>> +++ b/host/include/aarch64/host/aes-round.h
>> @@ -0,0 +1,204 @@
>> +/*
>> + * AArch64 specific aes acceleration.
>> + * SPDX-License-Identifier: GPL-2.0-or-later
>> + */
>> +
>> +#ifndef HOST_AES_ROUND_H
>> +#define HOST_AES_ROUND_H
>> +
>> +#include "host/cpuinfo.h"
>> +#include <arm_neon.h>
>> +
>> +#ifdef __ARM_FEATURE_AES
>> +# define HAVE_AES_ACCEL  true
>> +# define ATTR_AES_ACCEL
>> +#else
>> +# define HAVE_AES_ACCEL  likely(cpuinfo & CPUINFO_AES)
>> +# define ATTR_AES_ACCEL  __attribute__((target("+crypto")))
>> +#endif
>> +
>> +static inline uint8x16_t aes_accel_bswap(uint8x16_t x)
>> +{
>> +    /* No arm_neon.h primitive, and the compilers don't share builtins. */
> 
> vqtbl1q_u8() perhaps?

Ah, yes, thanks.


>> +static inline uint8x16_t aes_accel_aesmc(uint8x16_t d)
>> +{
>> +    asm(".arch_extension aes\n\t"
>> +        "aesmc %0.16b, %1.16b" : "=w"(d) : "w"(d));
> 
> 
> Most ARM cores fuse aese/aesmc into a single uop (with the associated
> performance boost) if the pattern is
> 
> aese x, y
> aesmc x,x
> 
> aesd x, y
> aesimc x,x
> 
> So it might make sense to use +w here at least, and use only a single
> register (which the compiler will likely do in any case, but still)
> 
> I would assume that the compiler cannot issue these separately based
> on the sequences below, but if it might, it may be worth it to emit
> the aese/aesmc together in a single asm() block

There could be shuffling.  It's low probability, but possible.

I really should move the builtin test to meson, as clang-16 fixes the builtin visibility 
issue.  I can see that gcc knows fusion of these pairs; I assume clang does as well, but I 
don't know the code base well enough to check.

I suppose it's going to be years until clang-16 can be assumed, as Debian bookworm is to 
be released this month with clang-14.  So it's probably worth spending a few more minutes 
on this now.


r~